2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
11 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
15 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
19 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
22 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23 * Copyright (C) IBM Corporation, 2009
24 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
26 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27 * enough at me, Linus for the original (flawed) idea, Matthew
28 * Kirkwood for proof-of-concept implementation.
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
33 * This program is free software; you can redistribute it and/or modify
34 * it under the terms of the GNU General Public License as published by
35 * the Free Software Foundation; either version 2 of the License, or
36 * (at your option) any later version.
38 * This program is distributed in the hope that it will be useful,
39 * but WITHOUT ANY WARRANTY; without even the implied warranty of
40 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
41 * GNU General Public License for more details.
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
47 #include <linux/slab.h>
48 #include <linux/poll.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64 #include <linux/hugetlb.h>
65 #include <linux/freezer.h>
66 #include <linux/bootmem.h>
68 #include <asm/futex.h>
70 #include "locking/rtmutex_common.h"
73 * READ this before attempting to hack on futexes!
75 * Basic futex operation and ordering guarantees
76 * =============================================
78 * The waiter reads the futex value in user space and calls
79 * futex_wait(). This function computes the hash bucket and acquires
80 * the hash bucket lock. After that it reads the futex user space value
81 * again and verifies that the data has not changed. If it has not changed
82 * it enqueues itself into the hash bucket, releases the hash bucket lock
85 * The waker side modifies the user space value of the futex and calls
86 * futex_wake(). This function computes the hash bucket and acquires the
87 * hash bucket lock. Then it looks for waiters on that futex in the hash
88 * bucket and wakes them.
90 * In futex wake up scenarios where no tasks are blocked on a futex, taking
91 * the hb spinlock can be avoided and simply return. In order for this
92 * optimization to work, ordering guarantees must exist so that the waiter
93 * being added to the list is acknowledged when the list is concurrently being
94 * checked by the waker, avoiding scenarios like the following:
98 * sys_futex(WAIT, futex, val);
99 * futex_wait(futex, val);
102 * sys_futex(WAKE, futex);
107 * lock(hash_bucket(futex));
109 * unlock(hash_bucket(futex));
112 * This would cause the waiter on CPU 0 to wait forever because it
113 * missed the transition of the user space value from val to newval
114 * and the waker did not find the waiter in the hash bucket queue.
116 * The correct serialization ensures that a waiter either observes
117 * the changed user space value before blocking or is woken by a
122 * sys_futex(WAIT, futex, val);
123 * futex_wait(futex, val);
126 * mb(); (A) <-- paired with -.
128 * lock(hash_bucket(futex)); |
132 * | sys_futex(WAKE, futex);
133 * | futex_wake(futex);
135 * `-------> mb(); (B)
138 * unlock(hash_bucket(futex));
139 * schedule(); if (waiters)
140 * lock(hash_bucket(futex));
141 * else wake_waiters(futex);
142 * waiters--; (b) unlock(hash_bucket(futex));
144 * Where (A) orders the waiters increment and the futex value read through
145 * atomic operations (see hb_waiters_inc) and where (B) orders the write
146 * to futex and the waiters read -- this is done by the barriers in
147 * get_futex_key_refs(), through either ihold or atomic_inc, depending on the
150 * This yields the following case (where X:=waiters, Y:=futex):
158 * Which guarantees that x==0 && y==0 is impossible; which translates back into
159 * the guarantee that we cannot both miss the futex variable change and the
162 * Note that a new waiter is accounted for in (a) even when it is possible that
163 * the wait call can return error, in which case we backtrack from it in (b).
164 * Refer to the comment in queue_lock().
166 * Similarly, in order to account for waiters being requeued on another
167 * address we always increment the waiters for the destination bucket before
168 * acquiring the lock. It then decrements them again after releasing it -
169 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
170 * will do the additional required waiter count housekeeping. This is done for
171 * double_lock_hb() and double_unlock_hb(), respectively.
174 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
175 int __read_mostly futex_cmpxchg_enabled
;
179 * Futex flags used to encode options to functions and preserve them across
182 #define FLAGS_SHARED 0x01
183 #define FLAGS_CLOCKRT 0x02
184 #define FLAGS_HAS_TIMEOUT 0x04
187 * Priority Inheritance state:
189 struct futex_pi_state
{
191 * list of 'owned' pi_state instances - these have to be
192 * cleaned up in do_exit() if the task exits prematurely:
194 struct list_head list
;
199 struct rt_mutex pi_mutex
;
201 struct task_struct
*owner
;
208 * struct futex_q - The hashed futex queue entry, one per waiting task
209 * @list: priority-sorted list of tasks waiting on this futex
210 * @task: the task waiting on the futex
211 * @lock_ptr: the hash bucket lock
212 * @key: the key the futex is hashed on
213 * @pi_state: optional priority inheritance state
214 * @rt_waiter: rt_waiter storage for use with requeue_pi
215 * @requeue_pi_key: the requeue_pi target futex key
216 * @bitset: bitset for the optional bitmasked wakeup
218 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
219 * we can wake only the relevant ones (hashed queues may be shared).
221 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
222 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
223 * The order of wakeup is always to make the first condition true, then
226 * PI futexes are typically woken before they are removed from the hash list via
227 * the rt_mutex code. See unqueue_me_pi().
230 struct plist_node list
;
232 struct task_struct
*task
;
233 spinlock_t
*lock_ptr
;
235 struct futex_pi_state
*pi_state
;
236 struct rt_mutex_waiter
*rt_waiter
;
237 union futex_key
*requeue_pi_key
;
241 static const struct futex_q futex_q_init
= {
242 /* list gets initialized in queue_me()*/
243 .key
= FUTEX_KEY_INIT
,
244 .bitset
= FUTEX_BITSET_MATCH_ANY
248 * Hash buckets are shared by all the futex_keys that hash to the same
249 * location. Each key may have multiple futex_q structures, one for each task
250 * waiting on a futex.
252 struct futex_hash_bucket
{
255 struct plist_head chain
;
256 } ____cacheline_aligned_in_smp
;
258 static unsigned long __read_mostly futex_hashsize
;
260 static struct futex_hash_bucket
*futex_queues
;
262 static inline void futex_get_mm(union futex_key
*key
)
264 atomic_inc(&key
->private.mm
->mm_count
);
266 * Ensure futex_get_mm() implies a full barrier such that
267 * get_futex_key() implies a full barrier. This is relied upon
268 * as full barrier (B), see the ordering comment above.
270 smp_mb__after_atomic();
274 * Reflects a new waiter being added to the waitqueue.
276 static inline void hb_waiters_inc(struct futex_hash_bucket
*hb
)
279 atomic_inc(&hb
->waiters
);
281 * Full barrier (A), see the ordering comment above.
283 smp_mb__after_atomic();
288 * Reflects a waiter being removed from the waitqueue by wakeup
291 static inline void hb_waiters_dec(struct futex_hash_bucket
*hb
)
294 atomic_dec(&hb
->waiters
);
298 static inline int hb_waiters_pending(struct futex_hash_bucket
*hb
)
301 return atomic_read(&hb
->waiters
);
308 * We hash on the keys returned from get_futex_key (see below).
310 static struct futex_hash_bucket
*hash_futex(union futex_key
*key
)
312 u32 hash
= jhash2((u32
*)&key
->both
.word
,
313 (sizeof(key
->both
.word
)+sizeof(key
->both
.ptr
))/4,
315 return &futex_queues
[hash
& (futex_hashsize
- 1)];
319 * Return 1 if two futex_keys are equal, 0 otherwise.
321 static inline int match_futex(union futex_key
*key1
, union futex_key
*key2
)
324 && key1
->both
.word
== key2
->both
.word
325 && key1
->both
.ptr
== key2
->both
.ptr
326 && key1
->both
.offset
== key2
->both
.offset
);
330 * Take a reference to the resource addressed by a key.
331 * Can be called while holding spinlocks.
334 static void get_futex_key_refs(union futex_key
*key
)
339 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
341 ihold(key
->shared
.inode
); /* implies MB (B) */
343 case FUT_OFF_MMSHARED
:
344 futex_get_mm(key
); /* implies MB (B) */
347 smp_mb(); /* explicit MB (B) */
352 * Drop a reference to the resource addressed by a key.
353 * The hash bucket spinlock must not be held.
355 static void drop_futex_key_refs(union futex_key
*key
)
357 if (!key
->both
.ptr
) {
358 /* If we're here then we tried to put a key we failed to get */
363 switch (key
->both
.offset
& (FUT_OFF_INODE
|FUT_OFF_MMSHARED
)) {
365 iput(key
->shared
.inode
);
367 case FUT_OFF_MMSHARED
:
368 mmdrop(key
->private.mm
);
374 * get_futex_key() - Get parameters which are the keys for a futex
375 * @uaddr: virtual address of the futex
376 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
377 * @key: address where result is stored.
378 * @rw: mapping needs to be read/write (values: VERIFY_READ,
381 * Return: a negative error code or 0
383 * The key words are stored in *key on success.
385 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
386 * offset_within_page). For private mappings, it's (uaddr, current->mm).
387 * We can usually work out the index without swapping in the page.
389 * lock_page() might sleep, the caller should not hold a spinlock.
392 get_futex_key(u32 __user
*uaddr
, int fshared
, union futex_key
*key
, int rw
)
394 unsigned long address
= (unsigned long)uaddr
;
395 struct mm_struct
*mm
= current
->mm
;
396 struct page
*page
, *page_head
;
400 * The futex address must be "naturally" aligned.
402 key
->both
.offset
= address
% PAGE_SIZE
;
403 if (unlikely((address
% sizeof(u32
)) != 0))
405 address
-= key
->both
.offset
;
407 if (unlikely(!access_ok(rw
, uaddr
, sizeof(u32
))))
411 * PROCESS_PRIVATE futexes are fast.
412 * As the mm cannot disappear under us and the 'key' only needs
413 * virtual address, we dont even have to find the underlying vma.
414 * Note : We do have to check 'uaddr' is a valid user address,
415 * but access_ok() should be faster than find_vma()
418 key
->private.mm
= mm
;
419 key
->private.address
= address
;
420 get_futex_key_refs(key
); /* implies MB (B) */
425 err
= get_user_pages_fast(address
, 1, 1, &page
);
427 * If write access is not required (eg. FUTEX_WAIT), try
428 * and get read-only access.
430 if (err
== -EFAULT
&& rw
== VERIFY_READ
) {
431 err
= get_user_pages_fast(address
, 1, 0, &page
);
439 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
441 if (unlikely(PageTail(page
))) {
443 /* serialize against __split_huge_page_splitting() */
445 if (likely(__get_user_pages_fast(address
, 1, !ro
, &page
) == 1)) {
446 page_head
= compound_head(page
);
448 * page_head is valid pointer but we must pin
449 * it before taking the PG_lock and/or
450 * PG_compound_lock. The moment we re-enable
451 * irqs __split_huge_page_splitting() can
452 * return and the head page can be freed from
453 * under us. We can't take the PG_lock and/or
454 * PG_compound_lock on a page that could be
455 * freed from under us.
457 if (page
!= page_head
) {
468 page_head
= compound_head(page
);
469 if (page
!= page_head
) {
475 lock_page(page_head
);
478 * If page_head->mapping is NULL, then it cannot be a PageAnon
479 * page; but it might be the ZERO_PAGE or in the gate area or
480 * in a special mapping (all cases which we are happy to fail);
481 * or it may have been a good file page when get_user_pages_fast
482 * found it, but truncated or holepunched or subjected to
483 * invalidate_complete_page2 before we got the page lock (also
484 * cases which we are happy to fail). And we hold a reference,
485 * so refcount care in invalidate_complete_page's remove_mapping
486 * prevents drop_caches from setting mapping to NULL beneath us.
488 * The case we do have to guard against is when memory pressure made
489 * shmem_writepage move it from filecache to swapcache beneath us:
490 * an unlikely race, but we do need to retry for page_head->mapping.
492 if (!page_head
->mapping
) {
493 int shmem_swizzled
= PageSwapCache(page_head
);
494 unlock_page(page_head
);
502 * Private mappings are handled in a simple way.
504 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
505 * it's a read-only handle, it's expected that futexes attach to
506 * the object not the particular process.
508 if (PageAnon(page_head
)) {
510 * A RO anonymous page will never change and thus doesn't make
511 * sense for futex operations.
518 key
->both
.offset
|= FUT_OFF_MMSHARED
; /* ref taken on mm */
519 key
->private.mm
= mm
;
520 key
->private.address
= address
;
522 key
->both
.offset
|= FUT_OFF_INODE
; /* inode-based key */
523 key
->shared
.inode
= page_head
->mapping
->host
;
524 key
->shared
.pgoff
= basepage_index(page
);
527 get_futex_key_refs(key
); /* implies MB (B) */
530 unlock_page(page_head
);
535 static inline void put_futex_key(union futex_key
*key
)
537 drop_futex_key_refs(key
);
541 * fault_in_user_writeable() - Fault in user address and verify RW access
542 * @uaddr: pointer to faulting user space address
544 * Slow path to fixup the fault we just took in the atomic write
547 * We have no generic implementation of a non-destructive write to the
548 * user address. We know that we faulted in the atomic pagefault
549 * disabled section so we can as well avoid the #PF overhead by
550 * calling get_user_pages() right away.
552 static int fault_in_user_writeable(u32 __user
*uaddr
)
554 struct mm_struct
*mm
= current
->mm
;
557 down_read(&mm
->mmap_sem
);
558 ret
= fixup_user_fault(current
, mm
, (unsigned long)uaddr
,
560 up_read(&mm
->mmap_sem
);
562 return ret
< 0 ? ret
: 0;
566 * futex_top_waiter() - Return the highest priority waiter on a futex
567 * @hb: the hash bucket the futex_q's reside in
568 * @key: the futex key (to distinguish it from other futex futex_q's)
570 * Must be called with the hb lock held.
572 static struct futex_q
*futex_top_waiter(struct futex_hash_bucket
*hb
,
573 union futex_key
*key
)
575 struct futex_q
*this;
577 plist_for_each_entry(this, &hb
->chain
, list
) {
578 if (match_futex(&this->key
, key
))
584 static int cmpxchg_futex_value_locked(u32
*curval
, u32 __user
*uaddr
,
585 u32 uval
, u32 newval
)
590 ret
= futex_atomic_cmpxchg_inatomic(curval
, uaddr
, uval
, newval
);
596 static int get_futex_value_locked(u32
*dest
, u32 __user
*from
)
601 ret
= __copy_from_user_inatomic(dest
, from
, sizeof(u32
));
604 return ret
? -EFAULT
: 0;
611 static int refill_pi_state_cache(void)
613 struct futex_pi_state
*pi_state
;
615 if (likely(current
->pi_state_cache
))
618 pi_state
= kzalloc(sizeof(*pi_state
), GFP_KERNEL
);
623 INIT_LIST_HEAD(&pi_state
->list
);
624 /* pi_mutex gets initialized later */
625 pi_state
->owner
= NULL
;
626 atomic_set(&pi_state
->refcount
, 1);
627 pi_state
->key
= FUTEX_KEY_INIT
;
629 current
->pi_state_cache
= pi_state
;
634 static struct futex_pi_state
* alloc_pi_state(void)
636 struct futex_pi_state
*pi_state
= current
->pi_state_cache
;
639 current
->pi_state_cache
= NULL
;
644 static void free_pi_state(struct futex_pi_state
*pi_state
)
646 if (!atomic_dec_and_test(&pi_state
->refcount
))
650 * If pi_state->owner is NULL, the owner is most probably dying
651 * and has cleaned up the pi_state already
653 if (pi_state
->owner
) {
654 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
655 list_del_init(&pi_state
->list
);
656 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
658 rt_mutex_proxy_unlock(&pi_state
->pi_mutex
, pi_state
->owner
);
661 if (current
->pi_state_cache
)
665 * pi_state->list is already empty.
666 * clear pi_state->owner.
667 * refcount is at 0 - put it back to 1.
669 pi_state
->owner
= NULL
;
670 atomic_set(&pi_state
->refcount
, 1);
671 current
->pi_state_cache
= pi_state
;
676 * Look up the task based on what TID userspace gave us.
679 static struct task_struct
* futex_find_get_task(pid_t pid
)
681 struct task_struct
*p
;
684 p
= find_task_by_vpid(pid
);
694 * This task is holding PI mutexes at exit time => bad.
695 * Kernel cleans up PI-state, but userspace is likely hosed.
696 * (Robust-futex cleanup is separate and might save the day for userspace.)
698 void exit_pi_state_list(struct task_struct
*curr
)
700 struct list_head
*next
, *head
= &curr
->pi_state_list
;
701 struct futex_pi_state
*pi_state
;
702 struct futex_hash_bucket
*hb
;
703 union futex_key key
= FUTEX_KEY_INIT
;
705 if (!futex_cmpxchg_enabled
)
708 * We are a ZOMBIE and nobody can enqueue itself on
709 * pi_state_list anymore, but we have to be careful
710 * versus waiters unqueueing themselves:
712 raw_spin_lock_irq(&curr
->pi_lock
);
713 while (!list_empty(head
)) {
716 pi_state
= list_entry(next
, struct futex_pi_state
, list
);
718 hb
= hash_futex(&key
);
719 raw_spin_unlock_irq(&curr
->pi_lock
);
721 spin_lock(&hb
->lock
);
723 raw_spin_lock_irq(&curr
->pi_lock
);
725 * We dropped the pi-lock, so re-check whether this
726 * task still owns the PI-state:
728 if (head
->next
!= next
) {
729 spin_unlock(&hb
->lock
);
733 WARN_ON(pi_state
->owner
!= curr
);
734 WARN_ON(list_empty(&pi_state
->list
));
735 list_del_init(&pi_state
->list
);
736 pi_state
->owner
= NULL
;
737 raw_spin_unlock_irq(&curr
->pi_lock
);
739 rt_mutex_unlock(&pi_state
->pi_mutex
);
741 spin_unlock(&hb
->lock
);
743 raw_spin_lock_irq(&curr
->pi_lock
);
745 raw_spin_unlock_irq(&curr
->pi_lock
);
749 * We need to check the following states:
751 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
753 * [1] NULL | --- | --- | 0 | 0/1 | Valid
754 * [2] NULL | --- | --- | >0 | 0/1 | Valid
756 * [3] Found | NULL | -- | Any | 0/1 | Invalid
758 * [4] Found | Found | NULL | 0 | 1 | Valid
759 * [5] Found | Found | NULL | >0 | 1 | Invalid
761 * [6] Found | Found | task | 0 | 1 | Valid
763 * [7] Found | Found | NULL | Any | 0 | Invalid
765 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
766 * [9] Found | Found | task | 0 | 0 | Invalid
767 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
769 * [1] Indicates that the kernel can acquire the futex atomically. We
770 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
772 * [2] Valid, if TID does not belong to a kernel thread. If no matching
773 * thread is found then it indicates that the owner TID has died.
775 * [3] Invalid. The waiter is queued on a non PI futex
777 * [4] Valid state after exit_robust_list(), which sets the user space
778 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
780 * [5] The user space value got manipulated between exit_robust_list()
781 * and exit_pi_state_list()
783 * [6] Valid state after exit_pi_state_list() which sets the new owner in
784 * the pi_state but cannot access the user space value.
786 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
788 * [8] Owner and user space value match
790 * [9] There is no transient state which sets the user space TID to 0
791 * except exit_robust_list(), but this is indicated by the
792 * FUTEX_OWNER_DIED bit. See [4]
794 * [10] There is no transient state which leaves owner and user space
799 * Validate that the existing waiter has a pi_state and sanity check
800 * the pi_state against the user space value. If correct, attach to
803 static int attach_to_pi_state(u32 uval
, struct futex_pi_state
*pi_state
,
804 struct futex_pi_state
**ps
)
806 pid_t pid
= uval
& FUTEX_TID_MASK
;
809 * Userspace might have messed up non-PI and PI futexes [3]
811 if (unlikely(!pi_state
))
814 WARN_ON(!atomic_read(&pi_state
->refcount
));
817 * Handle the owner died case:
819 if (uval
& FUTEX_OWNER_DIED
) {
821 * exit_pi_state_list sets owner to NULL and wakes the
822 * topmost waiter. The task which acquires the
823 * pi_state->rt_mutex will fixup owner.
825 if (!pi_state
->owner
) {
827 * No pi state owner, but the user space TID
828 * is not 0. Inconsistent state. [5]
833 * Take a ref on the state and return success. [4]
839 * If TID is 0, then either the dying owner has not
840 * yet executed exit_pi_state_list() or some waiter
841 * acquired the rtmutex in the pi state, but did not
842 * yet fixup the TID in user space.
844 * Take a ref on the state and return success. [6]
850 * If the owner died bit is not set, then the pi_state
851 * must have an owner. [7]
853 if (!pi_state
->owner
)
858 * Bail out if user space manipulated the futex value. If pi
859 * state exists then the owner TID must be the same as the
860 * user space TID. [9/10]
862 if (pid
!= task_pid_vnr(pi_state
->owner
))
865 atomic_inc(&pi_state
->refcount
);
871 * Lookup the task for the TID provided from user space and attach to
872 * it after doing proper sanity checks.
874 static int attach_to_pi_owner(u32 uval
, union futex_key
*key
,
875 struct futex_pi_state
**ps
)
877 pid_t pid
= uval
& FUTEX_TID_MASK
;
878 struct futex_pi_state
*pi_state
;
879 struct task_struct
*p
;
882 * We are the first waiter - try to look up the real owner and attach
883 * the new pi_state to it, but bail out when TID = 0 [1]
887 p
= futex_find_get_task(pid
);
897 * We need to look at the task state flags to figure out,
898 * whether the task is exiting. To protect against the do_exit
899 * change of the task flags, we do this protected by
902 raw_spin_lock_irq(&p
->pi_lock
);
903 if (unlikely(p
->flags
& PF_EXITING
)) {
905 * The task is on the way out. When PF_EXITPIDONE is
906 * set, we know that the task has finished the
909 int ret
= (p
->flags
& PF_EXITPIDONE
) ? -ESRCH
: -EAGAIN
;
911 raw_spin_unlock_irq(&p
->pi_lock
);
917 * No existing pi state. First waiter. [2]
919 pi_state
= alloc_pi_state();
922 * Initialize the pi_mutex in locked state and make @p
925 rt_mutex_init_proxy_locked(&pi_state
->pi_mutex
, p
);
927 /* Store the key for possible exit cleanups: */
928 pi_state
->key
= *key
;
930 WARN_ON(!list_empty(&pi_state
->list
));
931 list_add(&pi_state
->list
, &p
->pi_state_list
);
933 raw_spin_unlock_irq(&p
->pi_lock
);
942 static int lookup_pi_state(u32 uval
, struct futex_hash_bucket
*hb
,
943 union futex_key
*key
, struct futex_pi_state
**ps
)
945 struct futex_q
*match
= futex_top_waiter(hb
, key
);
948 * If there is a waiter on that futex, validate it and
949 * attach to the pi_state when the validation succeeds.
952 return attach_to_pi_state(uval
, match
->pi_state
, ps
);
955 * We are the first waiter - try to look up the owner based on
956 * @uval and attach to it.
958 return attach_to_pi_owner(uval
, key
, ps
);
961 static int lock_pi_update_atomic(u32 __user
*uaddr
, u32 uval
, u32 newval
)
963 u32
uninitialized_var(curval
);
965 if (unlikely(cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
)))
968 /*If user space value changed, let the caller retry */
969 return curval
!= uval
? -EAGAIN
: 0;
973 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
974 * @uaddr: the pi futex user address
975 * @hb: the pi futex hash bucket
976 * @key: the futex key associated with uaddr and hb
977 * @ps: the pi_state pointer where we store the result of the
979 * @task: the task to perform the atomic lock work for. This will
980 * be "current" except in the case of requeue pi.
981 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
985 * 1 - acquired the lock;
988 * The hb->lock and futex_key refs shall be held by the caller.
990 static int futex_lock_pi_atomic(u32 __user
*uaddr
, struct futex_hash_bucket
*hb
,
991 union futex_key
*key
,
992 struct futex_pi_state
**ps
,
993 struct task_struct
*task
, int set_waiters
)
995 u32 uval
, newval
, vpid
= task_pid_vnr(task
);
996 struct futex_q
*match
;
1000 * Read the user space value first so we can validate a few
1001 * things before proceeding further.
1003 if (get_futex_value_locked(&uval
, uaddr
))
1009 if ((unlikely((uval
& FUTEX_TID_MASK
) == vpid
)))
1013 * Lookup existing state first. If it exists, try to attach to
1016 match
= futex_top_waiter(hb
, key
);
1018 return attach_to_pi_state(uval
, match
->pi_state
, ps
);
1021 * No waiter and user TID is 0. We are here because the
1022 * waiters or the owner died bit is set or called from
1023 * requeue_cmp_pi or for whatever reason something took the
1026 if (!(uval
& FUTEX_TID_MASK
)) {
1028 * We take over the futex. No other waiters and the user space
1029 * TID is 0. We preserve the owner died bit.
1031 newval
= uval
& FUTEX_OWNER_DIED
;
1034 /* The futex requeue_pi code can enforce the waiters bit */
1036 newval
|= FUTEX_WAITERS
;
1038 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1039 /* If the take over worked, return 1 */
1040 return ret
< 0 ? ret
: 1;
1044 * First waiter. Set the waiters bit before attaching ourself to
1045 * the owner. If owner tries to unlock, it will be forced into
1046 * the kernel and blocked on hb->lock.
1048 newval
= uval
| FUTEX_WAITERS
;
1049 ret
= lock_pi_update_atomic(uaddr
, uval
, newval
);
1053 * If the update of the user space value succeeded, we try to
1054 * attach to the owner. If that fails, no harm done, we only
1055 * set the FUTEX_WAITERS bit in the user space variable.
1057 return attach_to_pi_owner(uval
, key
, ps
);
1061 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1062 * @q: The futex_q to unqueue
1064 * The q->lock_ptr must not be NULL and must be held by the caller.
1066 static void __unqueue_futex(struct futex_q
*q
)
1068 struct futex_hash_bucket
*hb
;
1070 if (WARN_ON_SMP(!q
->lock_ptr
|| !spin_is_locked(q
->lock_ptr
))
1071 || WARN_ON(plist_node_empty(&q
->list
)))
1074 hb
= container_of(q
->lock_ptr
, struct futex_hash_bucket
, lock
);
1075 plist_del(&q
->list
, &hb
->chain
);
1080 * The hash bucket lock must be held when this is called.
1081 * Afterwards, the futex_q must not be accessed.
1083 static void wake_futex(struct futex_q
*q
)
1085 struct task_struct
*p
= q
->task
;
1087 if (WARN(q
->pi_state
|| q
->rt_waiter
, "refusing to wake PI futex\n"))
1091 * We set q->lock_ptr = NULL _before_ we wake up the task. If
1092 * a non-futex wake up happens on another CPU then the task
1093 * might exit and p would dereference a non-existing task
1094 * struct. Prevent this by holding a reference on p across the
1101 * The waiting task can free the futex_q as soon as
1102 * q->lock_ptr = NULL is written, without taking any locks. A
1103 * memory barrier is required here to prevent the following
1104 * store to lock_ptr from getting ahead of the plist_del.
1109 wake_up_state(p
, TASK_NORMAL
);
1113 static int wake_futex_pi(u32 __user
*uaddr
, u32 uval
, struct futex_q
*this)
1115 struct task_struct
*new_owner
;
1116 struct futex_pi_state
*pi_state
= this->pi_state
;
1117 u32
uninitialized_var(curval
), newval
;
1124 * If current does not own the pi_state then the futex is
1125 * inconsistent and user space fiddled with the futex value.
1127 if (pi_state
->owner
!= current
)
1130 raw_spin_lock(&pi_state
->pi_mutex
.wait_lock
);
1131 new_owner
= rt_mutex_next_owner(&pi_state
->pi_mutex
);
1134 * It is possible that the next waiter (the one that brought
1135 * this owner to the kernel) timed out and is no longer
1136 * waiting on the lock.
1139 new_owner
= this->task
;
1142 * We pass it to the next owner. The WAITERS bit is always
1143 * kept enabled while there is PI state around. We cleanup the
1144 * owner died bit, because we are the owner.
1146 newval
= FUTEX_WAITERS
| task_pid_vnr(new_owner
);
1148 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
1150 else if (curval
!= uval
)
1153 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
1157 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
1158 WARN_ON(list_empty(&pi_state
->list
));
1159 list_del_init(&pi_state
->list
);
1160 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
1162 raw_spin_lock_irq(&new_owner
->pi_lock
);
1163 WARN_ON(!list_empty(&pi_state
->list
));
1164 list_add(&pi_state
->list
, &new_owner
->pi_state_list
);
1165 pi_state
->owner
= new_owner
;
1166 raw_spin_unlock_irq(&new_owner
->pi_lock
);
1168 raw_spin_unlock(&pi_state
->pi_mutex
.wait_lock
);
1169 rt_mutex_unlock(&pi_state
->pi_mutex
);
1175 * Express the locking dependencies for lockdep:
1178 double_lock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1181 spin_lock(&hb1
->lock
);
1183 spin_lock_nested(&hb2
->lock
, SINGLE_DEPTH_NESTING
);
1184 } else { /* hb1 > hb2 */
1185 spin_lock(&hb2
->lock
);
1186 spin_lock_nested(&hb1
->lock
, SINGLE_DEPTH_NESTING
);
1191 double_unlock_hb(struct futex_hash_bucket
*hb1
, struct futex_hash_bucket
*hb2
)
1193 spin_unlock(&hb1
->lock
);
1195 spin_unlock(&hb2
->lock
);
1199 * Wake up waiters matching bitset queued on this futex (uaddr).
1202 futex_wake(u32 __user
*uaddr
, unsigned int flags
, int nr_wake
, u32 bitset
)
1204 struct futex_hash_bucket
*hb
;
1205 struct futex_q
*this, *next
;
1206 union futex_key key
= FUTEX_KEY_INIT
;
1212 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_READ
);
1213 if (unlikely(ret
!= 0))
1216 hb
= hash_futex(&key
);
1218 /* Make sure we really have tasks to wakeup */
1219 if (!hb_waiters_pending(hb
))
1222 spin_lock(&hb
->lock
);
1224 plist_for_each_entry_safe(this, next
, &hb
->chain
, list
) {
1225 if (match_futex (&this->key
, &key
)) {
1226 if (this->pi_state
|| this->rt_waiter
) {
1231 /* Check if one of the bits is set in both bitsets */
1232 if (!(this->bitset
& bitset
))
1236 if (++ret
>= nr_wake
)
1241 spin_unlock(&hb
->lock
);
1243 put_futex_key(&key
);
1249 * Wake up all waiters hashed on the physical page that is mapped
1250 * to this virtual address:
1253 futex_wake_op(u32 __user
*uaddr1
, unsigned int flags
, u32 __user
*uaddr2
,
1254 int nr_wake
, int nr_wake2
, int op
)
1256 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1257 struct futex_hash_bucket
*hb1
, *hb2
;
1258 struct futex_q
*this, *next
;
1262 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1263 if (unlikely(ret
!= 0))
1265 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
1266 if (unlikely(ret
!= 0))
1269 hb1
= hash_futex(&key1
);
1270 hb2
= hash_futex(&key2
);
1273 double_lock_hb(hb1
, hb2
);
1274 op_ret
= futex_atomic_op_inuser(op
, uaddr2
);
1275 if (unlikely(op_ret
< 0)) {
1277 double_unlock_hb(hb1
, hb2
);
1281 * we don't get EFAULT from MMU faults if we don't have an MMU,
1282 * but we might get them from range checking
1288 if (unlikely(op_ret
!= -EFAULT
)) {
1293 ret
= fault_in_user_writeable(uaddr2
);
1297 if (!(flags
& FLAGS_SHARED
))
1300 put_futex_key(&key2
);
1301 put_futex_key(&key1
);
1305 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1306 if (match_futex (&this->key
, &key1
)) {
1307 if (this->pi_state
|| this->rt_waiter
) {
1312 if (++ret
>= nr_wake
)
1319 plist_for_each_entry_safe(this, next
, &hb2
->chain
, list
) {
1320 if (match_futex (&this->key
, &key2
)) {
1321 if (this->pi_state
|| this->rt_waiter
) {
1326 if (++op_ret
>= nr_wake2
)
1334 double_unlock_hb(hb1
, hb2
);
1336 put_futex_key(&key2
);
1338 put_futex_key(&key1
);
1344 * requeue_futex() - Requeue a futex_q from one hb to another
1345 * @q: the futex_q to requeue
1346 * @hb1: the source hash_bucket
1347 * @hb2: the target hash_bucket
1348 * @key2: the new key for the requeued futex_q
1351 void requeue_futex(struct futex_q
*q
, struct futex_hash_bucket
*hb1
,
1352 struct futex_hash_bucket
*hb2
, union futex_key
*key2
)
1356 * If key1 and key2 hash to the same bucket, no need to
1359 if (likely(&hb1
->chain
!= &hb2
->chain
)) {
1360 plist_del(&q
->list
, &hb1
->chain
);
1361 hb_waiters_dec(hb1
);
1362 plist_add(&q
->list
, &hb2
->chain
);
1363 hb_waiters_inc(hb2
);
1364 q
->lock_ptr
= &hb2
->lock
;
1366 get_futex_key_refs(key2
);
1371 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1373 * @key: the key of the requeue target futex
1374 * @hb: the hash_bucket of the requeue target futex
1376 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1377 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1378 * to the requeue target futex so the waiter can detect the wakeup on the right
1379 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1380 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1381 * to protect access to the pi_state to fixup the owner later. Must be called
1382 * with both q->lock_ptr and hb->lock held.
1385 void requeue_pi_wake_futex(struct futex_q
*q
, union futex_key
*key
,
1386 struct futex_hash_bucket
*hb
)
1388 get_futex_key_refs(key
);
1393 WARN_ON(!q
->rt_waiter
);
1394 q
->rt_waiter
= NULL
;
1396 q
->lock_ptr
= &hb
->lock
;
1398 wake_up_state(q
->task
, TASK_NORMAL
);
1402 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1403 * @pifutex: the user address of the to futex
1404 * @hb1: the from futex hash bucket, must be locked by the caller
1405 * @hb2: the to futex hash bucket, must be locked by the caller
1406 * @key1: the from futex key
1407 * @key2: the to futex key
1408 * @ps: address to store the pi_state pointer
1409 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1411 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1412 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1413 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1414 * hb1 and hb2 must be held by the caller.
1417 * 0 - failed to acquire the lock atomically;
1418 * >0 - acquired the lock, return value is vpid of the top_waiter
1421 static int futex_proxy_trylock_atomic(u32 __user
*pifutex
,
1422 struct futex_hash_bucket
*hb1
,
1423 struct futex_hash_bucket
*hb2
,
1424 union futex_key
*key1
, union futex_key
*key2
,
1425 struct futex_pi_state
**ps
, int set_waiters
)
1427 struct futex_q
*top_waiter
= NULL
;
1431 if (get_futex_value_locked(&curval
, pifutex
))
1435 * Find the top_waiter and determine if there are additional waiters.
1436 * If the caller intends to requeue more than 1 waiter to pifutex,
1437 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1438 * as we have means to handle the possible fault. If not, don't set
1439 * the bit unecessarily as it will force the subsequent unlock to enter
1442 top_waiter
= futex_top_waiter(hb1
, key1
);
1444 /* There are no waiters, nothing for us to do. */
1448 /* Ensure we requeue to the expected futex. */
1449 if (!match_futex(top_waiter
->requeue_pi_key
, key2
))
1453 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1454 * the contended case or if set_waiters is 1. The pi_state is returned
1455 * in ps in contended cases.
1457 vpid
= task_pid_vnr(top_waiter
->task
);
1458 ret
= futex_lock_pi_atomic(pifutex
, hb2
, key2
, ps
, top_waiter
->task
,
1461 requeue_pi_wake_futex(top_waiter
, key2
, hb2
);
1468 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1469 * @uaddr1: source futex user address
1470 * @flags: futex flags (FLAGS_SHARED, etc.)
1471 * @uaddr2: target futex user address
1472 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1473 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1474 * @cmpval: @uaddr1 expected value (or %NULL)
1475 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1476 * pi futex (pi to pi requeue is not supported)
1478 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1479 * uaddr2 atomically on behalf of the top waiter.
1482 * >=0 - on success, the number of tasks requeued or woken;
1485 static int futex_requeue(u32 __user
*uaddr1
, unsigned int flags
,
1486 u32 __user
*uaddr2
, int nr_wake
, int nr_requeue
,
1487 u32
*cmpval
, int requeue_pi
)
1489 union futex_key key1
= FUTEX_KEY_INIT
, key2
= FUTEX_KEY_INIT
;
1490 int drop_count
= 0, task_count
= 0, ret
;
1491 struct futex_pi_state
*pi_state
= NULL
;
1492 struct futex_hash_bucket
*hb1
, *hb2
;
1493 struct futex_q
*this, *next
;
1497 * Requeue PI only works on two distinct uaddrs. This
1498 * check is only valid for private futexes. See below.
1500 if (uaddr1
== uaddr2
)
1504 * requeue_pi requires a pi_state, try to allocate it now
1505 * without any locks in case it fails.
1507 if (refill_pi_state_cache())
1510 * requeue_pi must wake as many tasks as it can, up to nr_wake
1511 * + nr_requeue, since it acquires the rt_mutex prior to
1512 * returning to userspace, so as to not leave the rt_mutex with
1513 * waiters and no owner. However, second and third wake-ups
1514 * cannot be predicted as they involve race conditions with the
1515 * first wake and a fault while looking up the pi_state. Both
1516 * pthread_cond_signal() and pthread_cond_broadcast() should
1524 if (pi_state
!= NULL
) {
1526 * We will have to lookup the pi_state again, so free this one
1527 * to keep the accounting correct.
1529 free_pi_state(pi_state
);
1533 ret
= get_futex_key(uaddr1
, flags
& FLAGS_SHARED
, &key1
, VERIFY_READ
);
1534 if (unlikely(ret
!= 0))
1536 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
,
1537 requeue_pi
? VERIFY_WRITE
: VERIFY_READ
);
1538 if (unlikely(ret
!= 0))
1542 * The check above which compares uaddrs is not sufficient for
1543 * shared futexes. We need to compare the keys:
1545 if (requeue_pi
&& match_futex(&key1
, &key2
)) {
1550 hb1
= hash_futex(&key1
);
1551 hb2
= hash_futex(&key2
);
1554 hb_waiters_inc(hb2
);
1555 double_lock_hb(hb1
, hb2
);
1557 if (likely(cmpval
!= NULL
)) {
1560 ret
= get_futex_value_locked(&curval
, uaddr1
);
1562 if (unlikely(ret
)) {
1563 double_unlock_hb(hb1
, hb2
);
1564 hb_waiters_dec(hb2
);
1566 ret
= get_user(curval
, uaddr1
);
1570 if (!(flags
& FLAGS_SHARED
))
1573 put_futex_key(&key2
);
1574 put_futex_key(&key1
);
1577 if (curval
!= *cmpval
) {
1583 if (requeue_pi
&& (task_count
- nr_wake
< nr_requeue
)) {
1585 * Attempt to acquire uaddr2 and wake the top waiter. If we
1586 * intend to requeue waiters, force setting the FUTEX_WAITERS
1587 * bit. We force this here where we are able to easily handle
1588 * faults rather in the requeue loop below.
1590 ret
= futex_proxy_trylock_atomic(uaddr2
, hb1
, hb2
, &key1
,
1591 &key2
, &pi_state
, nr_requeue
);
1594 * At this point the top_waiter has either taken uaddr2 or is
1595 * waiting on it. If the former, then the pi_state will not
1596 * exist yet, look it up one more time to ensure we have a
1597 * reference to it. If the lock was taken, ret contains the
1598 * vpid of the top waiter task.
1605 * If we acquired the lock, then the user
1606 * space value of uaddr2 should be vpid. It
1607 * cannot be changed by the top waiter as it
1608 * is blocked on hb2 lock if it tries to do
1609 * so. If something fiddled with it behind our
1610 * back the pi state lookup might unearth
1611 * it. So we rather use the known value than
1612 * rereading and handing potential crap to
1615 ret
= lookup_pi_state(ret
, hb2
, &key2
, &pi_state
);
1622 double_unlock_hb(hb1
, hb2
);
1623 hb_waiters_dec(hb2
);
1624 put_futex_key(&key2
);
1625 put_futex_key(&key1
);
1626 ret
= fault_in_user_writeable(uaddr2
);
1632 * Two reasons for this:
1633 * - Owner is exiting and we just wait for the
1635 * - The user space value changed.
1637 double_unlock_hb(hb1
, hb2
);
1638 hb_waiters_dec(hb2
);
1639 put_futex_key(&key2
);
1640 put_futex_key(&key1
);
1648 plist_for_each_entry_safe(this, next
, &hb1
->chain
, list
) {
1649 if (task_count
- nr_wake
>= nr_requeue
)
1652 if (!match_futex(&this->key
, &key1
))
1656 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1657 * be paired with each other and no other futex ops.
1659 * We should never be requeueing a futex_q with a pi_state,
1660 * which is awaiting a futex_unlock_pi().
1662 if ((requeue_pi
&& !this->rt_waiter
) ||
1663 (!requeue_pi
&& this->rt_waiter
) ||
1670 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1671 * lock, we already woke the top_waiter. If not, it will be
1672 * woken by futex_unlock_pi().
1674 if (++task_count
<= nr_wake
&& !requeue_pi
) {
1679 /* Ensure we requeue to the expected futex for requeue_pi. */
1680 if (requeue_pi
&& !match_futex(this->requeue_pi_key
, &key2
)) {
1686 * Requeue nr_requeue waiters and possibly one more in the case
1687 * of requeue_pi if we couldn't acquire the lock atomically.
1690 /* Prepare the waiter to take the rt_mutex. */
1691 atomic_inc(&pi_state
->refcount
);
1692 this->pi_state
= pi_state
;
1693 ret
= rt_mutex_start_proxy_lock(&pi_state
->pi_mutex
,
1697 /* We got the lock. */
1698 requeue_pi_wake_futex(this, &key2
, hb2
);
1703 this->pi_state
= NULL
;
1704 free_pi_state(pi_state
);
1708 requeue_futex(this, hb1
, hb2
, &key2
);
1713 double_unlock_hb(hb1
, hb2
);
1714 hb_waiters_dec(hb2
);
1717 * drop_futex_key_refs() must be called outside the spinlocks. During
1718 * the requeue we moved futex_q's from the hash bucket at key1 to the
1719 * one at key2 and updated their key pointer. We no longer need to
1720 * hold the references to key1.
1722 while (--drop_count
>= 0)
1723 drop_futex_key_refs(&key1
);
1726 put_futex_key(&key2
);
1728 put_futex_key(&key1
);
1730 if (pi_state
!= NULL
)
1731 free_pi_state(pi_state
);
1732 return ret
? ret
: task_count
;
1735 /* The key must be already stored in q->key. */
1736 static inline struct futex_hash_bucket
*queue_lock(struct futex_q
*q
)
1737 __acquires(&hb
->lock
)
1739 struct futex_hash_bucket
*hb
;
1741 hb
= hash_futex(&q
->key
);
1744 * Increment the counter before taking the lock so that
1745 * a potential waker won't miss a to-be-slept task that is
1746 * waiting for the spinlock. This is safe as all queue_lock()
1747 * users end up calling queue_me(). Similarly, for housekeeping,
1748 * decrement the counter at queue_unlock() when some error has
1749 * occurred and we don't end up adding the task to the list.
1753 q
->lock_ptr
= &hb
->lock
;
1755 spin_lock(&hb
->lock
); /* implies MB (A) */
1760 queue_unlock(struct futex_hash_bucket
*hb
)
1761 __releases(&hb
->lock
)
1763 spin_unlock(&hb
->lock
);
1768 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1769 * @q: The futex_q to enqueue
1770 * @hb: The destination hash bucket
1772 * The hb->lock must be held by the caller, and is released here. A call to
1773 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1774 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1775 * or nothing if the unqueue is done as part of the wake process and the unqueue
1776 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1779 static inline void queue_me(struct futex_q
*q
, struct futex_hash_bucket
*hb
)
1780 __releases(&hb
->lock
)
1785 * The priority used to register this element is
1786 * - either the real thread-priority for the real-time threads
1787 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1788 * - or MAX_RT_PRIO for non-RT threads.
1789 * Thus, all RT-threads are woken first in priority order, and
1790 * the others are woken last, in FIFO order.
1792 prio
= min(current
->normal_prio
, MAX_RT_PRIO
);
1794 plist_node_init(&q
->list
, prio
);
1795 plist_add(&q
->list
, &hb
->chain
);
1797 spin_unlock(&hb
->lock
);
1801 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1802 * @q: The futex_q to unqueue
1804 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1805 * be paired with exactly one earlier call to queue_me().
1808 * 1 - if the futex_q was still queued (and we removed unqueued it);
1809 * 0 - if the futex_q was already removed by the waking thread
1811 static int unqueue_me(struct futex_q
*q
)
1813 spinlock_t
*lock_ptr
;
1816 /* In the common case we don't take the spinlock, which is nice. */
1818 lock_ptr
= q
->lock_ptr
;
1820 if (lock_ptr
!= NULL
) {
1821 spin_lock(lock_ptr
);
1823 * q->lock_ptr can change between reading it and
1824 * spin_lock(), causing us to take the wrong lock. This
1825 * corrects the race condition.
1827 * Reasoning goes like this: if we have the wrong lock,
1828 * q->lock_ptr must have changed (maybe several times)
1829 * between reading it and the spin_lock(). It can
1830 * change again after the spin_lock() but only if it was
1831 * already changed before the spin_lock(). It cannot,
1832 * however, change back to the original value. Therefore
1833 * we can detect whether we acquired the correct lock.
1835 if (unlikely(lock_ptr
!= q
->lock_ptr
)) {
1836 spin_unlock(lock_ptr
);
1841 BUG_ON(q
->pi_state
);
1843 spin_unlock(lock_ptr
);
1847 drop_futex_key_refs(&q
->key
);
1852 * PI futexes can not be requeued and must remove themself from the
1853 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1856 static void unqueue_me_pi(struct futex_q
*q
)
1857 __releases(q
->lock_ptr
)
1861 BUG_ON(!q
->pi_state
);
1862 free_pi_state(q
->pi_state
);
1865 spin_unlock(q
->lock_ptr
);
1869 * Fixup the pi_state owner with the new owner.
1871 * Must be called with hash bucket lock held and mm->sem held for non
1874 static int fixup_pi_state_owner(u32 __user
*uaddr
, struct futex_q
*q
,
1875 struct task_struct
*newowner
)
1877 u32 newtid
= task_pid_vnr(newowner
) | FUTEX_WAITERS
;
1878 struct futex_pi_state
*pi_state
= q
->pi_state
;
1879 struct task_struct
*oldowner
= pi_state
->owner
;
1880 u32 uval
, uninitialized_var(curval
), newval
;
1884 if (!pi_state
->owner
)
1885 newtid
|= FUTEX_OWNER_DIED
;
1888 * We are here either because we stole the rtmutex from the
1889 * previous highest priority waiter or we are the highest priority
1890 * waiter but failed to get the rtmutex the first time.
1891 * We have to replace the newowner TID in the user space variable.
1892 * This must be atomic as we have to preserve the owner died bit here.
1894 * Note: We write the user space value _before_ changing the pi_state
1895 * because we can fault here. Imagine swapped out pages or a fork
1896 * that marked all the anonymous memory readonly for cow.
1898 * Modifying pi_state _before_ the user space value would
1899 * leave the pi_state in an inconsistent state when we fault
1900 * here, because we need to drop the hash bucket lock to
1901 * handle the fault. This might be observed in the PID check
1902 * in lookup_pi_state.
1905 if (get_futex_value_locked(&uval
, uaddr
))
1909 newval
= (uval
& FUTEX_OWNER_DIED
) | newtid
;
1911 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, newval
))
1919 * We fixed up user space. Now we need to fix the pi_state
1922 if (pi_state
->owner
!= NULL
) {
1923 raw_spin_lock_irq(&pi_state
->owner
->pi_lock
);
1924 WARN_ON(list_empty(&pi_state
->list
));
1925 list_del_init(&pi_state
->list
);
1926 raw_spin_unlock_irq(&pi_state
->owner
->pi_lock
);
1929 pi_state
->owner
= newowner
;
1931 raw_spin_lock_irq(&newowner
->pi_lock
);
1932 WARN_ON(!list_empty(&pi_state
->list
));
1933 list_add(&pi_state
->list
, &newowner
->pi_state_list
);
1934 raw_spin_unlock_irq(&newowner
->pi_lock
);
1938 * To handle the page fault we need to drop the hash bucket
1939 * lock here. That gives the other task (either the highest priority
1940 * waiter itself or the task which stole the rtmutex) the
1941 * chance to try the fixup of the pi_state. So once we are
1942 * back from handling the fault we need to check the pi_state
1943 * after reacquiring the hash bucket lock and before trying to
1944 * do another fixup. When the fixup has been done already we
1948 spin_unlock(q
->lock_ptr
);
1950 ret
= fault_in_user_writeable(uaddr
);
1952 spin_lock(q
->lock_ptr
);
1955 * Check if someone else fixed it for us:
1957 if (pi_state
->owner
!= oldowner
)
1966 static long futex_wait_restart(struct restart_block
*restart
);
1969 * fixup_owner() - Post lock pi_state and corner case management
1970 * @uaddr: user address of the futex
1971 * @q: futex_q (contains pi_state and access to the rt_mutex)
1972 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1974 * After attempting to lock an rt_mutex, this function is called to cleanup
1975 * the pi_state owner as well as handle race conditions that may allow us to
1976 * acquire the lock. Must be called with the hb lock held.
1979 * 1 - success, lock taken;
1980 * 0 - success, lock not taken;
1981 * <0 - on error (-EFAULT)
1983 static int fixup_owner(u32 __user
*uaddr
, struct futex_q
*q
, int locked
)
1985 struct task_struct
*owner
;
1990 * Got the lock. We might not be the anticipated owner if we
1991 * did a lock-steal - fix up the PI-state in that case:
1993 if (q
->pi_state
->owner
!= current
)
1994 ret
= fixup_pi_state_owner(uaddr
, q
, current
);
1999 * Catch the rare case, where the lock was released when we were on the
2000 * way back before we locked the hash bucket.
2002 if (q
->pi_state
->owner
== current
) {
2004 * Try to get the rt_mutex now. This might fail as some other
2005 * task acquired the rt_mutex after we removed ourself from the
2006 * rt_mutex waiters list.
2008 if (rt_mutex_trylock(&q
->pi_state
->pi_mutex
)) {
2014 * pi_state is incorrect, some other task did a lock steal and
2015 * we returned due to timeout or signal without taking the
2016 * rt_mutex. Too late.
2018 raw_spin_lock(&q
->pi_state
->pi_mutex
.wait_lock
);
2019 owner
= rt_mutex_owner(&q
->pi_state
->pi_mutex
);
2021 owner
= rt_mutex_next_owner(&q
->pi_state
->pi_mutex
);
2022 raw_spin_unlock(&q
->pi_state
->pi_mutex
.wait_lock
);
2023 ret
= fixup_pi_state_owner(uaddr
, q
, owner
);
2028 * Paranoia check. If we did not take the lock, then we should not be
2029 * the owner of the rt_mutex.
2031 if (rt_mutex_owner(&q
->pi_state
->pi_mutex
) == current
)
2032 printk(KERN_ERR
"fixup_owner: ret = %d pi-mutex: %p "
2033 "pi-state %p\n", ret
,
2034 q
->pi_state
->pi_mutex
.owner
,
2035 q
->pi_state
->owner
);
2038 return ret
? ret
: locked
;
2042 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2043 * @hb: the futex hash bucket, must be locked by the caller
2044 * @q: the futex_q to queue up on
2045 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2047 static void futex_wait_queue_me(struct futex_hash_bucket
*hb
, struct futex_q
*q
,
2048 struct hrtimer_sleeper
*timeout
)
2051 * The task state is guaranteed to be set before another task can
2052 * wake it. set_current_state() is implemented using set_mb() and
2053 * queue_me() calls spin_unlock() upon completion, both serializing
2054 * access to the hash list and forcing another memory barrier.
2056 set_current_state(TASK_INTERRUPTIBLE
);
2061 hrtimer_start_expires(&timeout
->timer
, HRTIMER_MODE_ABS
);
2062 if (!hrtimer_active(&timeout
->timer
))
2063 timeout
->task
= NULL
;
2067 * If we have been removed from the hash list, then another task
2068 * has tried to wake us, and we can skip the call to schedule().
2070 if (likely(!plist_node_empty(&q
->list
))) {
2072 * If the timer has already expired, current will already be
2073 * flagged for rescheduling. Only call schedule if there
2074 * is no timeout, or if it has yet to expire.
2076 if (!timeout
|| timeout
->task
)
2077 freezable_schedule();
2079 __set_current_state(TASK_RUNNING
);
2083 * futex_wait_setup() - Prepare to wait on a futex
2084 * @uaddr: the futex userspace address
2085 * @val: the expected value
2086 * @flags: futex flags (FLAGS_SHARED, etc.)
2087 * @q: the associated futex_q
2088 * @hb: storage for hash_bucket pointer to be returned to caller
2090 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2091 * compare it with the expected value. Handle atomic faults internally.
2092 * Return with the hb lock held and a q.key reference on success, and unlocked
2093 * with no q.key reference on failure.
2096 * 0 - uaddr contains val and hb has been locked;
2097 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2099 static int futex_wait_setup(u32 __user
*uaddr
, u32 val
, unsigned int flags
,
2100 struct futex_q
*q
, struct futex_hash_bucket
**hb
)
2106 * Access the page AFTER the hash-bucket is locked.
2107 * Order is important:
2109 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2110 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2112 * The basic logical guarantee of a futex is that it blocks ONLY
2113 * if cond(var) is known to be true at the time of blocking, for
2114 * any cond. If we locked the hash-bucket after testing *uaddr, that
2115 * would open a race condition where we could block indefinitely with
2116 * cond(var) false, which would violate the guarantee.
2118 * On the other hand, we insert q and release the hash-bucket only
2119 * after testing *uaddr. This guarantees that futex_wait() will NOT
2120 * absorb a wakeup if *uaddr does not match the desired values
2121 * while the syscall executes.
2124 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
->key
, VERIFY_READ
);
2125 if (unlikely(ret
!= 0))
2129 *hb
= queue_lock(q
);
2131 ret
= get_futex_value_locked(&uval
, uaddr
);
2136 ret
= get_user(uval
, uaddr
);
2140 if (!(flags
& FLAGS_SHARED
))
2143 put_futex_key(&q
->key
);
2154 put_futex_key(&q
->key
);
2158 static int futex_wait(u32 __user
*uaddr
, unsigned int flags
, u32 val
,
2159 ktime_t
*abs_time
, u32 bitset
)
2161 struct hrtimer_sleeper timeout
, *to
= NULL
;
2162 struct restart_block
*restart
;
2163 struct futex_hash_bucket
*hb
;
2164 struct futex_q q
= futex_q_init
;
2174 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2175 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2177 hrtimer_init_sleeper(to
, current
);
2178 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2179 current
->timer_slack_ns
);
2184 * Prepare to wait on uaddr. On success, holds hb lock and increments
2187 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2191 /* queue_me and wait for wakeup, timeout, or a signal. */
2192 futex_wait_queue_me(hb
, &q
, to
);
2194 /* If we were woken (and unqueued), we succeeded, whatever. */
2196 /* unqueue_me() drops q.key ref */
2197 if (!unqueue_me(&q
))
2200 if (to
&& !to
->task
)
2204 * We expect signal_pending(current), but we might be the
2205 * victim of a spurious wakeup as well.
2207 if (!signal_pending(current
))
2214 restart
= ¤t_thread_info()->restart_block
;
2215 restart
->fn
= futex_wait_restart
;
2216 restart
->futex
.uaddr
= uaddr
;
2217 restart
->futex
.val
= val
;
2218 restart
->futex
.time
= abs_time
->tv64
;
2219 restart
->futex
.bitset
= bitset
;
2220 restart
->futex
.flags
= flags
| FLAGS_HAS_TIMEOUT
;
2222 ret
= -ERESTART_RESTARTBLOCK
;
2226 hrtimer_cancel(&to
->timer
);
2227 destroy_hrtimer_on_stack(&to
->timer
);
2233 static long futex_wait_restart(struct restart_block
*restart
)
2235 u32 __user
*uaddr
= restart
->futex
.uaddr
;
2236 ktime_t t
, *tp
= NULL
;
2238 if (restart
->futex
.flags
& FLAGS_HAS_TIMEOUT
) {
2239 t
.tv64
= restart
->futex
.time
;
2242 restart
->fn
= do_no_restart_syscall
;
2244 return (long)futex_wait(uaddr
, restart
->futex
.flags
,
2245 restart
->futex
.val
, tp
, restart
->futex
.bitset
);
2250 * Userspace tried a 0 -> TID atomic transition of the futex value
2251 * and failed. The kernel side here does the whole locking operation:
2252 * if there are waiters then it will block, it does PI, etc. (Due to
2253 * races the kernel might see a 0 value of the futex too.)
2255 static int futex_lock_pi(u32 __user
*uaddr
, unsigned int flags
, int detect
,
2256 ktime_t
*time
, int trylock
)
2258 struct hrtimer_sleeper timeout
, *to
= NULL
;
2259 struct futex_hash_bucket
*hb
;
2260 struct futex_q q
= futex_q_init
;
2263 if (refill_pi_state_cache())
2268 hrtimer_init_on_stack(&to
->timer
, CLOCK_REALTIME
,
2270 hrtimer_init_sleeper(to
, current
);
2271 hrtimer_set_expires(&to
->timer
, *time
);
2275 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &q
.key
, VERIFY_WRITE
);
2276 if (unlikely(ret
!= 0))
2280 hb
= queue_lock(&q
);
2282 ret
= futex_lock_pi_atomic(uaddr
, hb
, &q
.key
, &q
.pi_state
, current
, 0);
2283 if (unlikely(ret
)) {
2286 /* We got the lock. */
2288 goto out_unlock_put_key
;
2293 * Two reasons for this:
2294 * - Task is exiting and we just wait for the
2296 * - The user space value changed.
2299 put_futex_key(&q
.key
);
2303 goto out_unlock_put_key
;
2308 * Only actually queue now that the atomic ops are done:
2312 WARN_ON(!q
.pi_state
);
2314 * Block on the PI mutex:
2317 ret
= rt_mutex_timed_futex_lock(&q
.pi_state
->pi_mutex
, to
);
2319 ret
= rt_mutex_trylock(&q
.pi_state
->pi_mutex
);
2320 /* Fixup the trylock return value: */
2321 ret
= ret
? 0 : -EWOULDBLOCK
;
2324 spin_lock(q
.lock_ptr
);
2326 * Fixup the pi_state owner and possibly acquire the lock if we
2329 res
= fixup_owner(uaddr
, &q
, !ret
);
2331 * If fixup_owner() returned an error, proprogate that. If it acquired
2332 * the lock, clear our -ETIMEDOUT or -EINTR.
2335 ret
= (res
< 0) ? res
: 0;
2338 * If fixup_owner() faulted and was unable to handle the fault, unlock
2339 * it and return the fault to userspace.
2341 if (ret
&& (rt_mutex_owner(&q
.pi_state
->pi_mutex
) == current
))
2342 rt_mutex_unlock(&q
.pi_state
->pi_mutex
);
2344 /* Unqueue and drop the lock */
2353 put_futex_key(&q
.key
);
2356 destroy_hrtimer_on_stack(&to
->timer
);
2357 return ret
!= -EINTR
? ret
: -ERESTARTNOINTR
;
2362 ret
= fault_in_user_writeable(uaddr
);
2366 if (!(flags
& FLAGS_SHARED
))
2369 put_futex_key(&q
.key
);
2374 * Userspace attempted a TID -> 0 atomic transition, and failed.
2375 * This is the in-kernel slowpath: we look up the PI state (if any),
2376 * and do the rt-mutex unlock.
2378 static int futex_unlock_pi(u32 __user
*uaddr
, unsigned int flags
)
2380 u32
uninitialized_var(curval
), uval
, vpid
= task_pid_vnr(current
);
2381 union futex_key key
= FUTEX_KEY_INIT
;
2382 struct futex_hash_bucket
*hb
;
2383 struct futex_q
*match
;
2387 if (get_user(uval
, uaddr
))
2390 * We release only a lock we actually own:
2392 if ((uval
& FUTEX_TID_MASK
) != vpid
)
2395 ret
= get_futex_key(uaddr
, flags
& FLAGS_SHARED
, &key
, VERIFY_WRITE
);
2399 hb
= hash_futex(&key
);
2400 spin_lock(&hb
->lock
);
2403 * Check waiters first. We do not trust user space values at
2404 * all and we at least want to know if user space fiddled
2405 * with the futex value instead of blindly unlocking.
2407 match
= futex_top_waiter(hb
, &key
);
2409 ret
= wake_futex_pi(uaddr
, uval
, match
);
2411 * The atomic access to the futex value generated a
2412 * pagefault, so retry the user-access and the wakeup:
2420 * We have no kernel internal state, i.e. no waiters in the
2421 * kernel. Waiters which are about to queue themselves are stuck
2422 * on hb->lock. So we can safely ignore them. We do neither
2423 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2426 if (cmpxchg_futex_value_locked(&curval
, uaddr
, uval
, 0))
2430 * If uval has changed, let user space handle it.
2432 ret
= (curval
== uval
) ? 0 : -EAGAIN
;
2435 spin_unlock(&hb
->lock
);
2436 put_futex_key(&key
);
2440 spin_unlock(&hb
->lock
);
2441 put_futex_key(&key
);
2443 ret
= fault_in_user_writeable(uaddr
);
2451 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2452 * @hb: the hash_bucket futex_q was original enqueued on
2453 * @q: the futex_q woken while waiting to be requeued
2454 * @key2: the futex_key of the requeue target futex
2455 * @timeout: the timeout associated with the wait (NULL if none)
2457 * Detect if the task was woken on the initial futex as opposed to the requeue
2458 * target futex. If so, determine if it was a timeout or a signal that caused
2459 * the wakeup and return the appropriate error code to the caller. Must be
2460 * called with the hb lock held.
2463 * 0 = no early wakeup detected;
2464 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2467 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket
*hb
,
2468 struct futex_q
*q
, union futex_key
*key2
,
2469 struct hrtimer_sleeper
*timeout
)
2474 * With the hb lock held, we avoid races while we process the wakeup.
2475 * We only need to hold hb (and not hb2) to ensure atomicity as the
2476 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2477 * It can't be requeued from uaddr2 to something else since we don't
2478 * support a PI aware source futex for requeue.
2480 if (!match_futex(&q
->key
, key2
)) {
2481 WARN_ON(q
->lock_ptr
&& (&hb
->lock
!= q
->lock_ptr
));
2483 * We were woken prior to requeue by a timeout or a signal.
2484 * Unqueue the futex_q and determine which it was.
2486 plist_del(&q
->list
, &hb
->chain
);
2489 /* Handle spurious wakeups gracefully */
2491 if (timeout
&& !timeout
->task
)
2493 else if (signal_pending(current
))
2494 ret
= -ERESTARTNOINTR
;
2500 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2501 * @uaddr: the futex we initially wait on (non-pi)
2502 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2503 * the same type, no requeueing from private to shared, etc.
2504 * @val: the expected value of uaddr
2505 * @abs_time: absolute timeout
2506 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2507 * @uaddr2: the pi futex we will take prior to returning to user-space
2509 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2510 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2511 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2512 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2513 * without one, the pi logic would not know which task to boost/deboost, if
2514 * there was a need to.
2516 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2517 * via the following--
2518 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2519 * 2) wakeup on uaddr2 after a requeue
2523 * If 3, cleanup and return -ERESTARTNOINTR.
2525 * If 2, we may then block on trying to take the rt_mutex and return via:
2526 * 5) successful lock
2529 * 8) other lock acquisition failure
2531 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2533 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2539 static int futex_wait_requeue_pi(u32 __user
*uaddr
, unsigned int flags
,
2540 u32 val
, ktime_t
*abs_time
, u32 bitset
,
2543 struct hrtimer_sleeper timeout
, *to
= NULL
;
2544 struct rt_mutex_waiter rt_waiter
;
2545 struct rt_mutex
*pi_mutex
= NULL
;
2546 struct futex_hash_bucket
*hb
;
2547 union futex_key key2
= FUTEX_KEY_INIT
;
2548 struct futex_q q
= futex_q_init
;
2551 if (uaddr
== uaddr2
)
2559 hrtimer_init_on_stack(&to
->timer
, (flags
& FLAGS_CLOCKRT
) ?
2560 CLOCK_REALTIME
: CLOCK_MONOTONIC
,
2562 hrtimer_init_sleeper(to
, current
);
2563 hrtimer_set_expires_range_ns(&to
->timer
, *abs_time
,
2564 current
->timer_slack_ns
);
2568 * The waiter is allocated on our stack, manipulated by the requeue
2569 * code while we sleep on uaddr.
2571 debug_rt_mutex_init_waiter(&rt_waiter
);
2572 RB_CLEAR_NODE(&rt_waiter
.pi_tree_entry
);
2573 RB_CLEAR_NODE(&rt_waiter
.tree_entry
);
2574 rt_waiter
.task
= NULL
;
2576 ret
= get_futex_key(uaddr2
, flags
& FLAGS_SHARED
, &key2
, VERIFY_WRITE
);
2577 if (unlikely(ret
!= 0))
2581 q
.rt_waiter
= &rt_waiter
;
2582 q
.requeue_pi_key
= &key2
;
2585 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2588 ret
= futex_wait_setup(uaddr
, val
, flags
, &q
, &hb
);
2593 * The check above which compares uaddrs is not sufficient for
2594 * shared futexes. We need to compare the keys:
2596 if (match_futex(&q
.key
, &key2
)) {
2602 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2603 futex_wait_queue_me(hb
, &q
, to
);
2605 spin_lock(&hb
->lock
);
2606 ret
= handle_early_requeue_pi_wakeup(hb
, &q
, &key2
, to
);
2607 spin_unlock(&hb
->lock
);
2612 * In order for us to be here, we know our q.key == key2, and since
2613 * we took the hb->lock above, we also know that futex_requeue() has
2614 * completed and we no longer have to concern ourselves with a wakeup
2615 * race with the atomic proxy lock acquisition by the requeue code. The
2616 * futex_requeue dropped our key1 reference and incremented our key2
2620 /* Check if the requeue code acquired the second futex for us. */
2623 * Got the lock. We might not be the anticipated owner if we
2624 * did a lock-steal - fix up the PI-state in that case.
2626 if (q
.pi_state
&& (q
.pi_state
->owner
!= current
)) {
2627 spin_lock(q
.lock_ptr
);
2628 ret
= fixup_pi_state_owner(uaddr2
, &q
, current
);
2629 spin_unlock(q
.lock_ptr
);
2633 * We have been woken up by futex_unlock_pi(), a timeout, or a
2634 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2637 WARN_ON(!q
.pi_state
);
2638 pi_mutex
= &q
.pi_state
->pi_mutex
;
2639 ret
= rt_mutex_finish_proxy_lock(pi_mutex
, to
, &rt_waiter
);
2640 debug_rt_mutex_free_waiter(&rt_waiter
);
2642 spin_lock(q
.lock_ptr
);
2644 * Fixup the pi_state owner and possibly acquire the lock if we
2647 res
= fixup_owner(uaddr2
, &q
, !ret
);
2649 * If fixup_owner() returned an error, proprogate that. If it
2650 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2653 ret
= (res
< 0) ? res
: 0;
2655 /* Unqueue and drop the lock. */
2660 * If fixup_pi_state_owner() faulted and was unable to handle the
2661 * fault, unlock the rt_mutex and return the fault to userspace.
2663 if (ret
== -EFAULT
) {
2664 if (pi_mutex
&& rt_mutex_owner(pi_mutex
) == current
)
2665 rt_mutex_unlock(pi_mutex
);
2666 } else if (ret
== -EINTR
) {
2668 * We've already been requeued, but cannot restart by calling
2669 * futex_lock_pi() directly. We could restart this syscall, but
2670 * it would detect that the user space "val" changed and return
2671 * -EWOULDBLOCK. Save the overhead of the restart and return
2672 * -EWOULDBLOCK directly.
2678 put_futex_key(&q
.key
);
2680 put_futex_key(&key2
);
2684 hrtimer_cancel(&to
->timer
);
2685 destroy_hrtimer_on_stack(&to
->timer
);
2691 * Support for robust futexes: the kernel cleans up held futexes at
2694 * Implementation: user-space maintains a per-thread list of locks it
2695 * is holding. Upon do_exit(), the kernel carefully walks this list,
2696 * and marks all locks that are owned by this thread with the
2697 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2698 * always manipulated with the lock held, so the list is private and
2699 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2700 * field, to allow the kernel to clean up if the thread dies after
2701 * acquiring the lock, but just before it could have added itself to
2702 * the list. There can only be one such pending lock.
2706 * sys_set_robust_list() - Set the robust-futex list head of a task
2707 * @head: pointer to the list-head
2708 * @len: length of the list-head, as userspace expects
2710 SYSCALL_DEFINE2(set_robust_list
, struct robust_list_head __user
*, head
,
2713 if (!futex_cmpxchg_enabled
)
2716 * The kernel knows only one size for now:
2718 if (unlikely(len
!= sizeof(*head
)))
2721 current
->robust_list
= head
;
2727 * sys_get_robust_list() - Get the robust-futex list head of a task
2728 * @pid: pid of the process [zero for current task]
2729 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2730 * @len_ptr: pointer to a length field, the kernel fills in the header size
2732 SYSCALL_DEFINE3(get_robust_list
, int, pid
,
2733 struct robust_list_head __user
* __user
*, head_ptr
,
2734 size_t __user
*, len_ptr
)
2736 struct robust_list_head __user
*head
;
2738 struct task_struct
*p
;
2740 if (!futex_cmpxchg_enabled
)
2749 p
= find_task_by_vpid(pid
);
2755 if (!ptrace_may_access(p
, PTRACE_MODE_READ
))
2758 head
= p
->robust_list
;
2761 if (put_user(sizeof(*head
), len_ptr
))
2763 return put_user(head
, head_ptr
);
2772 * Process a futex-list entry, check whether it's owned by the
2773 * dying task, and do notification if so:
2775 int handle_futex_death(u32 __user
*uaddr
, struct task_struct
*curr
, int pi
)
2777 u32 uval
, uninitialized_var(nval
), mval
;
2780 if (get_user(uval
, uaddr
))
2783 if ((uval
& FUTEX_TID_MASK
) == task_pid_vnr(curr
)) {
2785 * Ok, this dying thread is truly holding a futex
2786 * of interest. Set the OWNER_DIED bit atomically
2787 * via cmpxchg, and if the value had FUTEX_WAITERS
2788 * set, wake up a waiter (if any). (We have to do a
2789 * futex_wake() even if OWNER_DIED is already set -
2790 * to handle the rare but possible case of recursive
2791 * thread-death.) The rest of the cleanup is done in
2794 mval
= (uval
& FUTEX_WAITERS
) | FUTEX_OWNER_DIED
;
2796 * We are not holding a lock here, but we want to have
2797 * the pagefault_disable/enable() protection because
2798 * we want to handle the fault gracefully. If the
2799 * access fails we try to fault in the futex with R/W
2800 * verification via get_user_pages. get_user() above
2801 * does not guarantee R/W access. If that fails we
2802 * give up and leave the futex locked.
2804 if (cmpxchg_futex_value_locked(&nval
, uaddr
, uval
, mval
)) {
2805 if (fault_in_user_writeable(uaddr
))
2813 * Wake robust non-PI futexes here. The wakeup of
2814 * PI futexes happens in exit_pi_state():
2816 if (!pi
&& (uval
& FUTEX_WAITERS
))
2817 futex_wake(uaddr
, 1, 1, FUTEX_BITSET_MATCH_ANY
);
2823 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2825 static inline int fetch_robust_entry(struct robust_list __user
**entry
,
2826 struct robust_list __user
* __user
*head
,
2829 unsigned long uentry
;
2831 if (get_user(uentry
, (unsigned long __user
*)head
))
2834 *entry
= (void __user
*)(uentry
& ~1UL);
2841 * Walk curr->robust_list (very carefully, it's a userspace list!)
2842 * and mark any locks found there dead, and notify any waiters.
2844 * We silently return on any sign of list-walking problem.
2846 void exit_robust_list(struct task_struct
*curr
)
2848 struct robust_list_head __user
*head
= curr
->robust_list
;
2849 struct robust_list __user
*entry
, *next_entry
, *pending
;
2850 unsigned int limit
= ROBUST_LIST_LIMIT
, pi
, pip
;
2851 unsigned int uninitialized_var(next_pi
);
2852 unsigned long futex_offset
;
2855 if (!futex_cmpxchg_enabled
)
2859 * Fetch the list head (which was registered earlier, via
2860 * sys_set_robust_list()):
2862 if (fetch_robust_entry(&entry
, &head
->list
.next
, &pi
))
2865 * Fetch the relative futex offset:
2867 if (get_user(futex_offset
, &head
->futex_offset
))
2870 * Fetch any possibly pending lock-add first, and handle it
2873 if (fetch_robust_entry(&pending
, &head
->list_op_pending
, &pip
))
2876 next_entry
= NULL
; /* avoid warning with gcc */
2877 while (entry
!= &head
->list
) {
2879 * Fetch the next entry in the list before calling
2880 * handle_futex_death:
2882 rc
= fetch_robust_entry(&next_entry
, &entry
->next
, &next_pi
);
2884 * A pending lock might already be on the list, so
2885 * don't process it twice:
2887 if (entry
!= pending
)
2888 if (handle_futex_death((void __user
*)entry
+ futex_offset
,
2896 * Avoid excessively long or circular lists:
2905 handle_futex_death((void __user
*)pending
+ futex_offset
,
2909 long do_futex(u32 __user
*uaddr
, int op
, u32 val
, ktime_t
*timeout
,
2910 u32 __user
*uaddr2
, u32 val2
, u32 val3
)
2912 int cmd
= op
& FUTEX_CMD_MASK
;
2913 unsigned int flags
= 0;
2915 if (!(op
& FUTEX_PRIVATE_FLAG
))
2916 flags
|= FLAGS_SHARED
;
2918 if (op
& FUTEX_CLOCK_REALTIME
) {
2919 flags
|= FLAGS_CLOCKRT
;
2920 if (cmd
!= FUTEX_WAIT_BITSET
&& cmd
!= FUTEX_WAIT_REQUEUE_PI
)
2926 case FUTEX_UNLOCK_PI
:
2927 case FUTEX_TRYLOCK_PI
:
2928 case FUTEX_WAIT_REQUEUE_PI
:
2929 case FUTEX_CMP_REQUEUE_PI
:
2930 if (!futex_cmpxchg_enabled
)
2936 val3
= FUTEX_BITSET_MATCH_ANY
;
2937 case FUTEX_WAIT_BITSET
:
2938 return futex_wait(uaddr
, flags
, val
, timeout
, val3
);
2940 val3
= FUTEX_BITSET_MATCH_ANY
;
2941 case FUTEX_WAKE_BITSET
:
2942 return futex_wake(uaddr
, flags
, val
, val3
);
2944 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, NULL
, 0);
2945 case FUTEX_CMP_REQUEUE
:
2946 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 0);
2948 return futex_wake_op(uaddr
, flags
, uaddr2
, val
, val2
, val3
);
2950 return futex_lock_pi(uaddr
, flags
, val
, timeout
, 0);
2951 case FUTEX_UNLOCK_PI
:
2952 return futex_unlock_pi(uaddr
, flags
);
2953 case FUTEX_TRYLOCK_PI
:
2954 return futex_lock_pi(uaddr
, flags
, 0, timeout
, 1);
2955 case FUTEX_WAIT_REQUEUE_PI
:
2956 val3
= FUTEX_BITSET_MATCH_ANY
;
2957 return futex_wait_requeue_pi(uaddr
, flags
, val
, timeout
, val3
,
2959 case FUTEX_CMP_REQUEUE_PI
:
2960 return futex_requeue(uaddr
, flags
, uaddr2
, val
, val2
, &val3
, 1);
2966 SYSCALL_DEFINE6(futex
, u32 __user
*, uaddr
, int, op
, u32
, val
,
2967 struct timespec __user
*, utime
, u32 __user
*, uaddr2
,
2971 ktime_t t
, *tp
= NULL
;
2973 int cmd
= op
& FUTEX_CMD_MASK
;
2975 if (utime
&& (cmd
== FUTEX_WAIT
|| cmd
== FUTEX_LOCK_PI
||
2976 cmd
== FUTEX_WAIT_BITSET
||
2977 cmd
== FUTEX_WAIT_REQUEUE_PI
)) {
2978 if (copy_from_user(&ts
, utime
, sizeof(ts
)) != 0)
2980 if (!timespec_valid(&ts
))
2983 t
= timespec_to_ktime(ts
);
2984 if (cmd
== FUTEX_WAIT
)
2985 t
= ktime_add_safe(ktime_get(), t
);
2989 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2990 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2992 if (cmd
== FUTEX_REQUEUE
|| cmd
== FUTEX_CMP_REQUEUE
||
2993 cmd
== FUTEX_CMP_REQUEUE_PI
|| cmd
== FUTEX_WAKE_OP
)
2994 val2
= (u32
) (unsigned long) utime
;
2996 return do_futex(uaddr
, op
, val
, tp
, uaddr2
, val2
, val3
);
2999 static void __init
futex_detect_cmpxchg(void)
3001 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3005 * This will fail and we want it. Some arch implementations do
3006 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3007 * functionality. We want to know that before we call in any
3008 * of the complex code paths. Also we want to prevent
3009 * registration of robust lists in that case. NULL is
3010 * guaranteed to fault and we get -EFAULT on functional
3011 * implementation, the non-functional ones will return
3014 if (cmpxchg_futex_value_locked(&curval
, NULL
, 0, 0) == -EFAULT
)
3015 futex_cmpxchg_enabled
= 1;
3019 static int __init
futex_init(void)
3021 unsigned int futex_shift
;
3024 #if CONFIG_BASE_SMALL
3025 futex_hashsize
= 16;
3027 futex_hashsize
= roundup_pow_of_two(256 * num_possible_cpus());
3030 futex_queues
= alloc_large_system_hash("futex", sizeof(*futex_queues
),
3032 futex_hashsize
< 256 ? HASH_SMALL
: 0,
3034 futex_hashsize
, futex_hashsize
);
3035 futex_hashsize
= 1UL << futex_shift
;
3037 futex_detect_cmpxchg();
3039 for (i
= 0; i
< futex_hashsize
; i
++) {
3040 atomic_set(&futex_queues
[i
].waiters
, 0);
3041 plist_head_init(&futex_queues
[i
].chain
);
3042 spin_lock_init(&futex_queues
[i
].lock
);
3047 __initcall(futex_init
);